Support Nemotron 3.5 Lightning

[FurtherAI]

Summary

• register the exact nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 checkpoint at revision b3caaabed0263651a17dc1f2d4ce97e794f76c44
• reuse the existing NemotronHHandler and already-qualified Nano Mamba-2, MoE, attention, context-parallel, LoRA, export, and native-vLLM paths because Lightning's 52-layer backbone geometry and schedule are identical
• deliberately disable Lightning's MTP configuration and exclude MTP tensors from HF validation; MTP training, adapters, export, and serving remain unsupported by design
• support Lightning's standard-Transformers config/checkpoint layout, built-in HF oracle, and model.layers LoRA namespace without changing Nano's remote-code backbone.layers contract
• make shared reduced fixtures understand serialized layers_block_type schedules and derive a missing scalar depth from that canonical schedule

Scope

This is an exact-checkpoint, same-architecture qualification. It does not change attention, FlexAttention, recurrent/Mamba-2, MoE, CP, or serving math. It does not alter train/inference thresholds, throughput floors, timeouts, lifecycle behavior, or compile settings. The throughput entry binds Lightning's exact workload fingerprint to the existing Nemotron-H profile without changing any floor.

Supported here: the BF16 backbone of NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 only. Lightning MTP, FP8/NVFP4 variants, Base/Super/Ultra variants, and multimodal checkpoints are not enabled.

Validation

• all-files prek hooks pass after rebasing onto current main
• focused provider, HF parity, padding, and workflow invariants: 96 passed, 1 intentional skip
• reduced HF parity: 61/61 checks; output MAPE 0.0017235%, worst gradient MAPE 0.321838% (FP32 0.5% gate)
• exact LoRA coverage/export: all 9 requested module families, no missing or unexpected families
• CP2/EP2 train/inference with native vLLM LoRA: 0.4566266% logprob MAPE (4% gate), 0.00068274 top-20 KL (0.002 gate), 98.4375% top-1 match
• FP32 packing invariance: all 4 fork/truncate/deep/repeated cases pass; worst logits MAPE 0.0057305%
• production-width six-layer CP2/EP2 trainability: one completed update, finite nonzero 22.933194 gradient norm, checkpoint publication, and generation from the exact updated native-vLLM adapter
• reduced throughput stage: 43,389.264 isolated physical tok/s, 43,311.030 matched E2E trainer-core tok/s, 0.998556 matched/isolated ratio, 41,555.020 E2E tok/s, 985.860 accepted tok/s, and 9.379706% useful MFU; W&B disabled

The complete default model-support workflow was intentionally not run for this same-architecture alias qualification. The optional correctness-sensitivity suite was also intentionally not run.